@ai-dossier/sched 0.19.0 → 0.21.0

package/README.md

@@ -82,7 +82,15 @@ where every mechanical supervision decision is code, not remembered prose:
    "waiting for it to finish" abandons the run with the subprocess still going; the
    instruction tells it to run such commands in the foreground and wait, or poll until
    they finish. `DEFAULT_REPORT_PROMPT_TEMPLATE` is excluded — it never spawns a long
-   command.
+   command. The prompt instruction alone was not enough (#591): agents kept arming the
+   `Monitor` tool to wait on a background command and ending their turn anyway, which the
+   engine can only see as an unverified exit. Every `claude`-family command template
+   (top-level `command` and each tier's own `commandTemplate`, #527) gets
+   `--disallowedTools Monitor` appended automatically — set `dispatch.disallowed_tools: []`
+   in `config.json` to opt out, or list your own tools to deny instead of the default
+   `["Monitor"]`. Matched on the binary's basename, so an absolute or wrapper path
+   (`/usr/local/bin/claude`) still gets it; never applied to a non-`claude` command or one
+   that already carries the flag itself, so an `opencode` tier is unaffected.
 2. **Completion verification (AC2)** — an agent exiting is never proof of completion.
    On exit, the unit completes only when ground truth confirms it: the issue's latest
    runstate milestone is `report done`, or GitHub says the issue is closed — except a
@@ -455,6 +463,8 @@ ready → executing(member i/N) ⟲ → validating → reviewing → shipping
 failure rails: executing → dissolving (a member self-reports blocked)
                validating → attributing → (fixing | evicting) → validating → dissolving
                validating → blocked (suite report unreadable, #562) → validating
+               executing → blocked (gate-inconclusive:<cap>, #583) → executing
+                 (`sched resume --batch <id>` re-runs the gate; nothing requeued/reverted)
 ```
 
 - **One shared worktree/branch per batch**, claimed once by a deterministic (no LLM)
@@ -473,7 +483,12 @@ failure rails: executing → dissolving (a member self-reports blocked)
   kept on `BatchEntry.ranges` for eviction. An incremental gate (`ai-dossier cap run
   typecheck.run` / `test.focused`, when the repo has a manifest) runs after each member
   before advancing — a second, independent check that the member's self-reported "done"
-  is real.
+  is real. Three-way outcome policy (#583): `task-failed` evicts the member (same rail
+  as a self-reported block); `automation-broken`/`capability-unavailable` — the gate
+  itself couldn't reach a verdict — block the batch instead of silently proceeding
+  (`gate-inconclusive:<cap>`, `member_gates`/`blocked_reason` on `BatchEntry`, surfaced
+  in `sched status`); `sched resume --batch <id>` re-runs the gate later to resolve the
+  block once the capability is fixed.
 - **The batch's single slot is claimed FRESH for each live step** (a member, the tail
   agent, the report agent, a bounded fix agent) — never held across a wait. The aggregate
   suite itself runs with NO slot claimed at all (deterministic engine work, not an LLM
@@ -504,10 +519,21 @@ batch-setup had pool integration). Config schema moves to 1.3.0: `dispatch` gain
 `member_prompt`, `batch_tail_prompt` and `batch_report_prompt` (the three new agent
 prompt templates).
 
+Schema 1.11.0 (#583): `BatchEntry` gains `member_gates` (most recent incremental-gate
+result per member, keyed by issue number as a string — `{capability, outcome,
+output_tail, at}`) and `blocked_reason` (why the batch is `blocked` — persisted so
+`sched status` can show it; previously `blockBatch` only journaled/posted the reason,
+never stored it on the entry, so this also retroactively covers the #562 case). 1.10.0
+states migrate on load: no gate has ever produced a non-`ok` verdict, and no batch has
+ever been blocked, under them, so `{}`/`null` is the exact backfill, not a guess.
+
 New journal events: `batch-setup-done`, `batch-setup-failed`, `member-advanced`,
 `batch-warmup-done`, `batch-warmup-failed` (#561 — the cold-path warm step only; a pool
-claim emits neither). Member/tail/report/fix-agent spawn, progress, completion and park
-events reuse the existing unit-generic names (`assigned`/`spawned`/`unit-failed`/
+claim emits neither). `gate-inconclusive` (#583 — the incremental gate came back
+`automation-broken`/`capability-unavailable` rather than a definite `ok`/`task-failed`;
+sits alongside `batch-blocked` as the per-member analogue of the aggregate suite's
+"block, don't dissolve" precedent). Member/tail/report/fix-agent spawn, progress,
+completion and park events reuse the existing unit-generic names (`assigned`/`spawned`/`unit-failed`/
 `external-advance`/`pr-parked`/`merge-accepted`/`report-dispatched`/`teardown-done`/
 `teardown-failed`) with `unit = batch:<id>`.
 
@@ -624,10 +650,12 @@ import {
                          //   the issue pass — loads/saves state itself, holds no lock
                          //   across the call
   type BatchDispatchDeps, // inject store/journal/groundTruth/spawnDeps/exec/runSuite/
-                         //   runCapability(optional)/fsExists(optional)
+                         //   runCapability(optional, returns CapabilityGateResult)/fsExists(optional)
   type BatchTickResult,  // spawned/completed/parked/mergeAccepted/failed (batch:<id> ids)
                          //   + blocked (issue numbers, dissolve-requeued)
   type CapOutcome,       // ok | task-failed | automation-broken | capability-unavailable
+  type CapabilityGateResult, // {outcome: CapOutcome, outputTail?, reason?} — runCapability's return shape (#583)
+  resumeBlockedGate,     // #583: sched resume --batch <id> — re-run the gate that blocked a batch
   buildMemberPrompt, buildBatchTailPrompt, buildBatchReportPrompt, // #523 prompt builders
   DEFAULT_MEMBER_PROMPT_TEMPLATE, DEFAULT_BATCH_TAIL_PROMPT_TEMPLATE,
   DEFAULT_BATCH_REPORT_PROMPT_TEMPLATE,
@@ -656,7 +684,7 @@ telemetry" below.
 ├── config.json    # durable intent: max_slots, stall_timeout_ms, reconcile_interval_ms,
 │                  # pr_poll_interval_ms, dispatch (incl. report_prompt,
 │                  # phase_stall_timeout_ms, fence_takeover_timeout_ms, tiers — #527,
-│                  # suite_command — #562), auto_upgrade — #537,
+│                  # suite_command — #562, disallowed_tools — #591), auto_upgrade — #537,
 │                  # dissolve_policy — #563
 ├── events.jsonl   # append-only event journal (the operator's flight recorder)
 ├── runs/          # per-unit agent output logs (issue-<n>.log)
@@ -677,6 +705,14 @@ every place a dispatch's exit is first detected: the dead-pid rail, the
 external-advance rail, a stall-timeout kill, and a dependents-blocked kill), sourced
 from the agent's `modelUsage` map — never blended with the top-level `usage` block,
 the fix for a ~43% fabricated-saving discrepancy the two blocks were found to produce.
+`recordDispatchRunLog`/`recordMemberRunLog` also return the last tool the dispatch called
+(`parseLastToolUse`, `@ai-dossier/core`, #591), when the log yielded one — the exit itself
+attributes to a concrete cause (e.g. `Monitor`) without opening the transcript. It rides the
+non-terminal `verify-incomplete` event on every unverified exit and, once the escalation
+ladder is exhausted, the terminal `unit-failed` (`agent-exited-unverified` /
+`unverified-exit-at-strongest-tier`) as `last_tool`; a stall-timeout kill carries it too. Only
+the dead-pid detection rail and the stall kill record a fresh log slice in the same tick —
+a slot already `exited`/`verifying` when reconciled again has none to attribute.
 
 The dispatch log (`runs/<unit>.log`) is per-UNIT and opened in append mode
 (`createSpawnDeps`), so a redispatched unit's second agent writes its output AFTER the

package/dist/batch-dispatch.d.ts

@@ -67,15 +67,8 @@ import type { SchedStore } from './persist';
 import type { ExecFn } from './project';
 import { type SuiteResult } from './recovery';
 import { type FsExists } from './teardown';
-import type { SchedConfig } from './types';
-/**
- * The four `ai-dossier cap run` outcomes (docs/reference/capabilities.md):
- * `ok` = the capability ran and passed; `task-failed` = it ran and the TASK
- * itself failed (trust the result); `automation-broken` = do not trust the
- * machinery (missing tool, timeout, bad manifest); `capability-unavailable` =
- * no manifest / no such id / `lifecycle: shadow` — no fast path here.
- */
-export type CapOutcome = 'ok' | 'task-failed' | 'automation-broken' | 'capability-unavailable';
+import type { CapabilityGateResult, SchedConfig } from './types';
+export type { CapOutcome } from './types';
 /** Everything batch dispatch needs from the outside world. */
 export interface BatchDispatchDeps {
     store: SchedStore;
@@ -100,19 +93,24 @@ export interface BatchDispatchDeps {
     /**
      * Runs one `ai-dossier cap run <capabilityId>` in a batch worktree. Two call
      * sites, different degrade contracts:
-     * - the per-member incremental gate (#523 AC2 — "typecheck + focused tests
-     *   via `cap run test.focused` when available"): degrade-not-crash — without
-     *   this hook, or on `automation-broken`/`capability-unavailable`, the gate
-     *   is skipped (the member's own `slot-cycle` run already attempted this
-     *   fast path before ever posting `review done`, so a repo with no manifest
-     *   loses nothing but the engine's independent re-check).
+     * - the per-member incremental gate (#523 AC2, revised #583): without this
+     *   hook, the gate is skipped entirely (the member's own `slot-cycle` run
+     *   already attempted this fast path before ever posting `review done`, so
+     *   a repo with no manifest loses nothing but the engine's independent
+     *   re-check). With the hook: `ok` advances, `task-failed` evicts,
+     *   `automation-broken`/`capability-unavailable` BLOCK the batch
+     *   (`gate-inconclusive`) rather than silently proceeding — #583 found the
+     *   old "proceed on anything but task-failed" policy let a script's own
+     *   "I could not run this" signal (a non-zero exit reporting it never
+     *   really tested anything) masquerade as either a pass or a real failure.
      * - batch-setup's `worktree.prepare` warm step (#561, `warmColdBatchWorktree`):
-     *   `undefined`/`capability-unavailable`/`automation-broken` fall through to
-     *   package-manager detection, but a declared-and-`task-failed` capability
-     *   hard-fails the whole batch setup — a repo that owns its warm-up should
-     *   never be silently second-guessed by a fallback underneath it.
+     *   unaffected by #583's gate policy change — `undefined`/
+     *   `capability-unavailable`/`automation-broken` still fall through to
+     *   package-manager detection, and a declared-and-`task-failed` capability
+     *   still hard-fails the whole batch setup (a repo that owns its warm-up
+     *   should never be silently second-guessed by a fallback underneath it).
      */
-    runCapability?: (worktree: string, capabilityId: string) => CapOutcome;
+    runCapability?: (worktree: string, capabilityId: string) => CapabilityGateResult;
     fsExists?: FsExists;
     /**
      * Home directory for `~/.dossier/runs.jsonl` (#564) — mirrors `EngineDeps.homeDir`.
@@ -130,6 +128,33 @@ export interface BatchTickResult {
     /** Issue numbers requeued full-cycle by a dissolve — matches `TickResult.blocked`'s shape. */
     blocked: number[];
 }
+/**
+ * `sched resume --batch <id>` (#583 AC4): an operator-triggered, synchronous
+ * one-shot recheck of a batch blocked on `gate-inconclusive:<capabilityId>`
+ * — re-runs exactly that capability against the current member and resolves
+ * the block:
+ *
+ * - still `automation-broken`/`capability-unavailable` → stays `blocked`,
+ *   no state change (the capability still isn't fixed).
+ * - `task-failed` → the member really is broken; evict via the same rail
+ *   the live gate uses.
+ * - `ok` → the member really was fine; complete it via the same rail the
+ *   live gate uses.
+ *
+ * Deliberately NOT a generic engine tick: `runBatchTick`'s own "executing
+ * with no live slot" wedge-recovery path (`spawnMemberContinuation`,
+ * unconditional) exists for a dispatch that never happened — reusing it here
+ * would redispatch a fresh agent for a member whose work is already
+ * committed and reviewed. This function transitions `blocked → executing`
+ * and immediately, synchronously, calls the same completion/eviction
+ * functions the live gate calls — no intervening tick ever sees the batch
+ * `executing` with nothing in flight.
+ */
+export declare function resumeBlockedGate(deps: BatchDispatchDeps, config: SchedConfig, dispatch: ResolvedDispatch, batchId: string, now: Date): {
+    outcome: 'still-blocked' | 'evicted' | 'completed';
+    capability: string;
+    detail?: string;
+};
 /**
  * One batch reconcile+refill pass. Called from `engine.ts`'s `tick()` after
  * the issue-level pass — this pass never claims a slot the issue pass already

package/dist/batch-dispatch.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"batch-dispatch.d.ts","sourceRoot":"","sources":["../src/batch-dispatch.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA6DG;AAiBH,OAAO,EASL,KAAK,gBAAgB,EAErB,KAAK,SAAS,EACf,MAAM,YAAY,CAAC;AACpB,OAAO,EACL,KAAK,WAAW,EAOjB,MAAM,eAAe,CAAC;AACvB,OAAO,EAAE,KAAK,OAAO,EAAa,MAAM,WAAW,CAAC;AACpD,OAAO,KAAK,EAAE,UAAU,EAAE,MAAM,WAAW,CAAC;AAC5C,OAAO,KAAK,EAAE,MAAM,EAAE,MAAM,WAAW,CAAC;AAExC,OAAO,EAUL,KAAK,WAAW,EACjB,MAAM,YAAY,CAAC;AAYpB,OAAO,EAAE,KAAK,QAAQ,EAA2D,MAAM,YAAY,CAAC;AACpG,OAAO,KAAK,EAKV,WAAW,EAIZ,MAAM,SAAS,CAAC;AAGjB;;;;;;GAMG;AACH,MAAM,MAAM,UAAU,GAAG,IAAI,GAAG,aAAa,GAAG,mBAAmB,GAAG,wBAAwB,CAAC;AAE/F,8DAA8D;AAC9D,MAAM,WAAW,iBAAiB;IAChC,KAAK,EAAE,UAAU,CAAC;IAClB,OAAO,EAAE,OAAO,CAAC;IACjB,WAAW,EAAE,WAAW,CAAC;IACzB,SAAS,EAAE,SAAS,CAAC;IACrB,GAAG,EAAE,MAAM,IAAI,CAAC;IAChB,uFAAuF;IACvF,OAAO,EAAE,MAAM,CAAC;IAChB,0FAA0F;IAC1F,IAAI,EAAE,MAAM,CAAC;IACb;;;;;;OAMG;IACH,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,sGAAsG;IACtG,QAAQ,EAAE,CAAC,QAAQ,EAAE,MAAM,KAAK,WAAW,CAAC;IAC5C;;;;;;;;;;;;;;OAcG;IACH,aAAa,CAAC,EAAE,CAAC,QAAQ,EAAE,MAAM,EAAE,YAAY,EAAE,MAAM,KAAK,UAAU,CAAC;IACvE,QAAQ,CAAC,EAAE,QAAQ,CAAC;IACpB;;;OAGG;IACH,OAAO,CAAC,EAAE,MAAM,CAAC;CAClB;AAED,8FAA8F;AAC9F,MAAM,WAAW,eAAe;IAC9B,OAAO,EAAE,MAAM,EAAE,CAAC;IAClB,SAAS,EAAE,MAAM,EAAE,CAAC;IACpB,MAAM,EAAE,MAAM,EAAE,CAAC;IACjB,aAAa,EAAE,MAAM,EAAE,CAAC;IACxB,MAAM,EAAE,MAAM,EAAE,CAAC;IACjB,8FAA8F;IAC9F,OAAO,EAAE,MAAM,EAAE,CAAC;CACnB;AA2jDD;;;;;;;;;;;;;GAaG;AACH,wBAAgB,YAAY,CAC1B,IAAI,EAAE,iBAAiB,EACvB,MAAM,EAAE,WAAW,EACnB,QAAQ,EAAE,gBAAgB,GACzB,eAAe,CAsFjB"}
+{"version":3,"file":"batch-dispatch.d.ts","sourceRoot":"","sources":["../src/batch-dispatch.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA6DG;AAkBH,OAAO,EAUL,KAAK,gBAAgB,EAErB,KAAK,SAAS,EACf,MAAM,YAAY,CAAC;AACpB,OAAO,EACL,KAAK,WAAW,EAOjB,MAAM,eAAe,CAAC;AACvB,OAAO,EAAE,KAAK,OAAO,EAAa,MAAM,WAAW,CAAC;AACpD,OAAO,KAAK,EAAE,UAAU,EAAE,MAAM,WAAW,CAAC;AAC5C,OAAO,KAAK,EAAE,MAAM,EAAE,MAAM,WAAW,CAAC;AAExC,OAAO,EAUL,KAAK,WAAW,EACjB,MAAM,YAAY,CAAC;AAYpB,OAAO,EAAE,KAAK,QAAQ,EAA2D,MAAM,YAAY,CAAC;AACpG,OAAO,KAAK,EAGV,oBAAoB,EAIpB,WAAW,EAIZ,MAAM,SAAS,CAAC;AAMjB,YAAY,EAAE,UAAU,EAAE,MAAM,SAAS,CAAC;AAE1C,8DAA8D;AAC9D,MAAM,WAAW,iBAAiB;IAChC,KAAK,EAAE,UAAU,CAAC;IAClB,OAAO,EAAE,OAAO,CAAC;IACjB,WAAW,EAAE,WAAW,CAAC;IACzB,SAAS,EAAE,SAAS,CAAC;IACrB,GAAG,EAAE,MAAM,IAAI,CAAC;IAChB,uFAAuF;IACvF,OAAO,EAAE,MAAM,CAAC;IAChB,0FAA0F;IAC1F,IAAI,EAAE,MAAM,CAAC;IACb;;;;;;OAMG;IACH,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,sGAAsG;IACtG,QAAQ,EAAE,CAAC,QAAQ,EAAE,MAAM,KAAK,WAAW,CAAC;IAC5C;;;;;;;;;;;;;;;;;;;OAmBG;IACH,aAAa,CAAC,EAAE,CAAC,QAAQ,EAAE,MAAM,EAAE,YAAY,EAAE,MAAM,KAAK,oBAAoB,CAAC;IACjF,QAAQ,CAAC,EAAE,QAAQ,CAAC;IACpB;;;OAGG;IACH,OAAO,CAAC,EAAE,MAAM,CAAC;CAClB;AAED,8FAA8F;AAC9F,MAAM,WAAW,eAAe;IAC9B,OAAO,EAAE,MAAM,EAAE,CAAC;IAClB,SAAS,EAAE,MAAM,EAAE,CAAC;IACpB,MAAM,EAAE,MAAM,EAAE,CAAC;IACjB,aAAa,EAAE,MAAM,EAAE,CAAC;IACxB,MAAM,EAAE,MAAM,EAAE,CAAC;IACjB,8FAA8F;IAC9F,OAAO,EAAE,MAAM,EAAE,CAAC;CACnB;AAywCD;;;;;;;;;;;;;;;;;;;;;GAqBG;AACH,wBAAgB,iBAAiB,CAC/B,IAAI,EAAE,iBAAiB,EACvB,MAAM,EAAE,WAAW,EACnB,QAAQ,EAAE,gBAAgB,EAC1B,OAAO,EAAE,MAAM,EACf,GAAG,EAAE,IAAI,GACR;IAAE,OAAO,EAAE,eAAe,GAAG,SAAS,GAAG,WAAW,CAAC;IAAC,UAAU,EAAE,MAAM,CAAC;IAAC,MAAM,CAAC,EAAE,MAAM,CAAA;CAAE,CA+E7F;AAofD;;;;;;;;;;;;;GAaG;AACH,wBAAgB,YAAY,CAC1B,IAAI,EAAE,iBAAiB,EACvB,MAAM,EAAE,WAAW,EACnB,QAAQ,EAAE,gBAAgB,GACzB,eAAe,CAsFjB"}

package/dist/batch-dispatch.js

@@ -95,9 +95,11 @@ var __importStar = (this && this.__importStar) || (function () {
     };
 })();
 Object.defineProperty(exports, "__esModule", { value: true });
+exports.resumeBlockedGate = resumeBlockedGate;
 exports.runBatchTick = runBatchTick;
 const fs = __importStar(require("node:fs"));
 const path = __importStar(require("node:path"));
+const core_1 = require("@ai-dossier/core");
 const worktree_pool_1 = require("@ai-dossier/worktree-pool");
 const attribution_1 = require("./attribution");
 const dispatch_1 = require("./dispatch");
@@ -140,6 +142,65 @@ function sanitizeUntrustedText(value) {
 function journalEvent(deps, event, unitId, extra = {}) {
     deps.journal.append((0, journal_1.unitEvent)(event, unitId, extra), deps.now());
 }
+/**
+ * Last ~500 bytes of a gate's output tail, falling back to the envelope's
+ * `reason` when no subprocess ran (#583 AC1/AC3 review: `capability-unavailable`
+ * and a failed assumption probe carry no output_tail — `reason` is the only
+ * explanation available for those) — journal details stay compact; the full
+ * tail lives in the per-gate log file. UTF-8-safe (never splits a multi-byte
+ * character) — chars would risk exactly that, hence `Buffer`, matching
+ * `cli/src/capability.ts`'s `truncateTailBytes`.
+ */
+function gateDetailExcerpt(outputTail, reason) {
+    const text = outputTail || reason;
+    if (!text)
+        return undefined;
+    const buf = Buffer.from(text, 'utf-8');
+    return buf.length > 500 ? buf.subarray(buf.length - 500).toString('utf-8') : text;
+}
+/**
+ * Best-effort per-gate diagnostic log (#583 AC1) — mirrors `appendCapLog`'s
+ * never-crash contract: a log-write failure must not interrupt the gate
+ * decision itself.
+ */
+function writeGateLog(deps, batchId, capabilityId, issue, outputTail) {
+    if (!outputTail)
+        return;
+    try {
+        const logPath = (0, dispatch_1.batchGateLogPath)(deps.store.runsDir, batchId, capabilityId, issue);
+        fs.mkdirSync(path.dirname(logPath), { recursive: true, mode: 0o700 });
+        fs.writeFileSync(logPath, outputTail, { mode: 0o600 });
+    }
+    catch {
+        // diagnostic only — never fail the gate decision over a log write
+    }
+}
+/**
+ * Persist the incremental gate's most recent verdict for a member on
+ * `BatchEntry.member_gates` (#583 AC4) — shared by the live gate
+ * (`reconcileMemberSlot`) and `sched resume --batch`'s recheck
+ * (`resumeBlockedGate`), so a resumed batch's status never shows a stale
+ * outcome from before the recheck ran.
+ */
+function recordMemberGate(deps, batchId, memberIssue, gate, now) {
+    deps.store.withLock((s) => {
+        const b = (0, state_1.findBatch)(s, batchId);
+        return {
+            state: (0, state_1.patchBatch)(s, batchId, {
+                member_gates: {
+                    ...(b?.member_gates ?? {}),
+                    [String(memberIssue)]: {
+                        capability: gate.id,
+                        outcome: gate.outcome,
+                        output_tail: gate.outputTail ?? null,
+                        at: now.toISOString(),
+                    },
+                },
+            }, now),
+            result: undefined,
+        };
+    });
+}
 function slotFor(state, batchId) {
     return state.slots.find((s) => s.unit === unit(batchId));
 }
@@ -288,7 +349,7 @@ poolNote = '') {
         }), now);
         return { ok: true };
     };
-    const capOutcome = deps.runCapability?.(worktree, 'worktree.prepare');
+    const capOutcome = deps.runCapability?.(worktree, 'worktree.prepare')?.outcome;
     if (capOutcome === 'ok')
         return done('cap:worktree.prepare');
     if (capOutcome === 'task-failed')
@@ -932,6 +993,187 @@ function evictMemberAndContinue(deps, config, dispatch, batchId, batch, memberIs
     }
     advanceMemberOrValidate(deps, config, dispatch, batchId, batch.members.length, batch.executing_member, memberIssue, now, result);
 }
+/**
+ * The incremental gate (#523 AC2, revised #583): typecheck + focused tests
+ * via `cap run`, when the repo has a manifest for them — a second,
+ * independent check that the member's own self-reported "done" is real,
+ * matching this codebase's "never trust a claimed completion" ethos
+ * (AC2/#464's `isVerifiedComplete`). Three-way policy: `task-failed` evicts
+ * the member directly, same rail as a self-reported block (RFC F.1) — no
+ * aggregate suite has run yet, so there is nothing to attribute.
+ * `automation-broken`/`capability-unavailable` — the gate itself could not
+ * reach a verdict — BLOCK the batch instead of silently proceeding (#583: a
+ * script that legitimately could not run its suite must not be read as
+ * either a pass or a real failure). Only both `ok` falls through.
+ *
+ * Returns `true` when the gate already decided the member's fate (evicted or
+ * blocked, both of which `return` from the caller); `false` when no hook is
+ * configured or both checks came back `ok`, meaning the caller should treat
+ * the member as genuinely complete.
+ */
+function runIncrementalGate(deps, config, dispatch, batchId, batch, memberIssue, now, result) {
+    if (batch.worktree === null || !deps.runCapability)
+        return false;
+    const worktree = batch.worktree;
+    const runCapability = deps.runCapability;
+    const gateResults = ['typecheck.run', 'test.focused'].map((id) => ({
+        id,
+        ...runCapability(worktree, id),
+    }));
+    const gateFailure = gateResults.find((r) => r.outcome === 'task-failed');
+    const gateInconclusive = gateResults.find((r) => r.outcome === 'automation-broken' || r.outcome === 'capability-unavailable');
+    const worstGate = gateFailure ?? gateInconclusive;
+    if (worstGate) {
+        recordMemberGate(deps, batchId, memberIssue, worstGate, now);
+    }
+    if (gateFailure) {
+        const reason = `incremental-gate-failed:${gateFailure.id}`;
+        writeGateLog(deps, batchId, gateFailure.id, memberIssue, gateFailure.outputTail);
+        const excerpt = gateDetailExcerpt(gateFailure.outputTail, gateFailure.reason);
+        journalEvent(deps, 'unit-failed', unit(batchId), {
+            issue: memberIssue,
+            reason,
+            detail: excerpt
+                ? `cap run ${gateFailure.id} reported task-failed after member review done: ${excerpt}`
+                : `cap run ${gateFailure.id} reported task-failed after member review done`,
+        });
+        deps.store.withLock((s) => ({ state: releaseSlot(s, batchId, now), result: undefined }));
+        evictMemberAndContinue(deps, config, dispatch, batchId, batch, memberIssue, reason, now, result);
+        return true;
+    }
+    if (gateInconclusive) {
+        const reason = `gate-inconclusive:${gateInconclusive.id}`;
+        writeGateLog(deps, batchId, gateInconclusive.id, memberIssue, gateInconclusive.outputTail);
+        const excerpt = gateDetailExcerpt(gateInconclusive.outputTail, gateInconclusive.reason);
+        journalEvent(deps, 'gate-inconclusive', unit(batchId), {
+            issue: memberIssue,
+            reason,
+            detail: excerpt
+                ? `cap run ${gateInconclusive.id} reported ${gateInconclusive.outcome} after member review done: ${excerpt}`
+                : `cap run ${gateInconclusive.id} reported ${gateInconclusive.outcome} after member review done`,
+        });
+        deps.store.withLock((s) => ({ state: releaseSlot(s, batchId, now), result: undefined }));
+        const stateNow = deps.store.load();
+        const rDeps = recoveryDeps(deps, config, batch, now);
+        const blocked = (0, recovery_1.blockBatch)(stateNow, batchId, { reason, milestonePhase: 'batch-review' }, rDeps);
+        deps.store.withLock((s) => ({
+            state: applyBatchAndIssues(s, blocked.state, batchId, []),
+            result: undefined,
+        }));
+        result.failed.push(unit(batchId));
+        return true;
+    }
+    return false;
+}
+/**
+ * The member is verifiably done — self-reported complete AND (when checked)
+ * the incremental gate agrees. Recompute the member's commit range, mark it
+ * validated, and advance the batch. Shared by `reconcileMemberSlot`'s
+ * immediate "both gate checks ok" fallthrough and `resumeBlockedGate` (#583)
+ * — a member confirmed complete via a delayed `sched resume --batch` recheck
+ * gets exactly the same treatment as one confirmed complete on the first try.
+ * `releaseSlot` is idempotent (a no-op once the slot is already idle), so
+ * this is safe to call whether or not the caller already released it.
+ */
+function completeMemberGate(deps, config, dispatch, batchId, batch, memberIssue, now, result) {
+    // The commit-range recompute (`git log`) is a blocking subprocess call — it
+    // must run OUTSIDE the lock, like every other exec in this module; the
+    // result then lands as a pure data patch under the lock (Convention review:
+    // `recordRanges` used to run `git log` INSIDE the withLock mutator, which
+    // is exactly what `engine.ts`'s own "a slow git call never holds the lock"
+    // invariant exists to prevent).
+    const ranges = (0, attribution_1.memberRanges)(boundaryCommits(deps, batch));
+    deps.store.withLock((s) => {
+        let n = releaseSlot(s, batchId, now);
+        n = (0, state_1.patchBatch)(n, batchId, { ranges }, now);
+        n = advanceMemberToValidated(n, memberIssue, now);
+        return { state: n, result: undefined };
+    });
+    result.completed.push(unit(batchId));
+    advanceMemberOrValidate(deps, config, dispatch, batchId, batch.members.length, batch.executing_member, memberIssue, now, result);
+}
+/**
+ * `sched resume --batch <id>` (#583 AC4): an operator-triggered, synchronous
+ * one-shot recheck of a batch blocked on `gate-inconclusive:<capabilityId>`
+ * — re-runs exactly that capability against the current member and resolves
+ * the block:
+ *
+ * - still `automation-broken`/`capability-unavailable` → stays `blocked`,
+ *   no state change (the capability still isn't fixed).
+ * - `task-failed` → the member really is broken; evict via the same rail
+ *   the live gate uses.
+ * - `ok` → the member really was fine; complete it via the same rail the
+ *   live gate uses.
+ *
+ * Deliberately NOT a generic engine tick: `runBatchTick`'s own "executing
+ * with no live slot" wedge-recovery path (`spawnMemberContinuation`,
+ * unconditional) exists for a dispatch that never happened — reusing it here
+ * would redispatch a fresh agent for a member whose work is already
+ * committed and reviewed. This function transitions `blocked → executing`
+ * and immediately, synchronously, calls the same completion/eviction
+ * functions the live gate calls — no intervening tick ever sees the batch
+ * `executing` with nothing in flight.
+ */
+function resumeBlockedGate(deps, config, dispatch, batchId, now) {
+    const state = deps.store.load();
+    const batch = (0, state_1.findBatch)(state, batchId);
+    if (!batch)
+        throw new types_1.SchedNotFoundError(`Batch not found: ${batchId}`);
+    if (batch.status !== 'blocked' || !batch.blocked_reason?.startsWith('gate-inconclusive:')) {
+        throw new types_1.IllegalTransitionError('batch', batch.status, 'executing');
+    }
+    const capabilityId = batch.blocked_reason.slice('gate-inconclusive:'.length);
+    const memberIssue = batch.members[batch.executing_member - 1];
+    if (batch.worktree === null || !deps.runCapability || memberIssue === undefined) {
+        throw new types_1.SchedNotFoundError(`Batch ${batchId} has no worktree/current member/gate hook to recheck`);
+    }
+    const recheck = deps.runCapability(batch.worktree, capabilityId);
+    const result = emptyResult();
+    const excerpt = gateDetailExcerpt(recheck.outputTail, recheck.reason);
+    recordMemberGate(deps, batchId, memberIssue, { id: capabilityId, ...recheck }, now);
+    if (recheck.outcome === 'automation-broken' || recheck.outcome === 'capability-unavailable') {
+        // Still inconclusive: leave an audit trail (log + journal) exactly like
+        // the live gate does, even though the batch stays `blocked` — otherwise
+        // a repeated `sched resume --batch` leaves no record of when it was last
+        // checked or what it said (#583 review).
+        writeGateLog(deps, batchId, capabilityId, memberIssue, recheck.outputTail);
+        journalEvent(deps, 'gate-inconclusive', unit(batchId), {
+            issue: memberIssue,
+            reason: `gate-inconclusive:${capabilityId}`,
+            detail: excerpt
+                ? `sched resume --batch: cap run ${capabilityId} still reports ${recheck.outcome} on recheck: ${excerpt}`
+                : `sched resume --batch: cap run ${capabilityId} still reports ${recheck.outcome} on recheck`,
+        });
+        return {
+            outcome: 'still-blocked',
+            capability: capabilityId,
+            detail: excerpt,
+        };
+    }
+    deps.store.withLock((s) => ({
+        state: (0, state_1.patchBatch)((0, state_1.transitionBatch)(s, batchId, 'executing', {}, now), batchId, { blocked_reason: null }, now),
+        result: undefined,
+    }));
+    if (recheck.outcome === 'task-failed') {
+        const reason = `incremental-gate-failed:${capabilityId}`;
+        writeGateLog(deps, batchId, capabilityId, memberIssue, recheck.outputTail);
+        journalEvent(deps, 'unit-failed', unit(batchId), {
+            issue: memberIssue,
+            reason,
+            detail: excerpt
+                ? `sched resume --batch: cap run ${capabilityId} reported task-failed on recheck: ${excerpt}`
+                : `sched resume --batch: cap run ${capabilityId} reported task-failed on recheck`,
+        });
+        evictMemberAndContinue(deps, config, dispatch, batchId, batch, memberIssue, reason, now, result);
+        return { outcome: 'evicted', capability: capabilityId, detail: excerpt };
+    }
+    journalEvent(deps, 'external-advance', unit(batchId), {
+        issue: memberIssue,
+        detail: `sched resume --batch: cap run ${capabilityId} reported ok on recheck`,
+    });
+    completeMemberGate(deps, config, dispatch, batchId, batch, memberIssue, now, result);
+    return { outcome: 'completed', capability: capabilityId };
+}
 /**
  * Record one member dispatch's tokens/cost to `runs.jsonl` (#564) — the
  * `batch-dispatch.ts` analogue of `engine.ts`'s `recordDispatchRunLog`.
@@ -957,7 +1199,7 @@ function recordMemberRunLog(deps, dispatch, state, batchId, batch, memberIssue,
             reason: slot.spawned_at === null ? 'never-spawned' : `already-recorded-${slot.status}`,
             slot: slot.id,
         });
-        return;
+        return null;
     }
     const tier = (0, state_1.findEntry)(state, memberIssue)?.tier ?? 'mid';
     const { cmd, model } = (0, dispatch_1.resolveTierSpawn)(dispatch, tier, memberIssue);
@@ -976,6 +1218,9 @@ function recordMemberRunLog(deps, dispatch, state, batchId, batch, memberIssue,
         tier,
     });
     (0, run_log_1.finalizeRunLogEntry)(runEntry, logContent, deps.homeDir, (event, extra) => journalEvent(deps, event, unit(batchId), extra), { issue: memberIssue, log: logFile });
+    // #591: the last tool this member dispatch called — attributes an
+    // `agent-exited-unverified` failure to a concrete cause without opening the transcript.
+    return (0, core_1.parseLastToolUse)(logContent);
 }
 function reconcileMemberSlot(deps, config, dispatch, batchId, slot, now, result) {
     const state0 = deps.store.load();
@@ -1016,46 +1261,15 @@ function reconcileMemberSlot(deps, config, dispatch, batchId, slot, now, result)
         // incremental gate below decides — record its telemetry once here (#564)
         // rather than at each of this branch's two later exit points.
         recordMemberRunLog(deps, dispatch, state0, batchId, batch, memberIssue, slot, now);
-        // Incremental gate (#523 AC2): typecheck + focused tests via `cap run`,
-        // when the repo has a manifest for them — a second, independent check
-        // that the member's own self-reported "done" is real, matching this
-        // codebase's "never trust a claimed completion" ethos (AC2/#464's
-        // `isVerifiedComplete`). `task-failed` evicts the member directly, same
-        // rail as a self-reported block (RFC F.1) — no aggregate suite has run
-        // yet, so there is nothing to attribute.
-        if (batch.worktree !== null && deps.runCapability) {
-            const gateFailure = ['typecheck.run', 'test.focused']
-                .map((id) => ({ id, outcome: deps.runCapability?.(batch.worktree, id) }))
-                .find((r) => r.outcome === 'task-failed');
-            if (gateFailure) {
-                const reason = `incremental-gate-failed:${gateFailure.id}`;
-                journalEvent(deps, 'unit-failed', unit(batchId), {
-                    issue: memberIssue,
-                    reason,
-                    detail: `cap run ${gateFailure.id} reported task-failed after member review done`,
-                });
-                deps.store.withLock((s) => ({ state: releaseSlot(s, batchId, now), result: undefined }));
-                evictMemberAndContinue(deps, config, dispatch, batchId, batch, memberIssue, reason, now, result);
-                return;
-            }
-            // `ok` / `automation-broken` / `capability-unavailable` all proceed —
-            // only a definite task failure blocks a member here.
+        // Incremental gate (#523 AC2, revised #583) — see `runIncrementalGate`'s
+        // own doc comment for the three-way policy. A `true` return means the
+        // gate already decided the member's fate (evicted or blocked) and
+        // returned early; `false` means both checks were `ok` and the member is
+        // genuinely done.
+        if (runIncrementalGate(deps, config, dispatch, batchId, batch, memberIssue, now, result)) {
+            return;
         }
-        // The commit-range recompute (`git log`) is a blocking subprocess call —
-        // it must run OUTSIDE the lock, like every other exec in this module;
-        // the result then lands as a pure data patch under the lock (Convention
-        // review: `recordRanges` used to run `git log` INSIDE the withLock
-        // mutator, which is exactly what `engine.ts`'s own "a slow git call never
-        // holds the lock" invariant exists to prevent).
-        const ranges = (0, attribution_1.memberRanges)(boundaryCommits(deps, batch));
-        deps.store.withLock((s) => {
-            let n = releaseSlot(s, batchId, now);
-            n = (0, state_1.patchBatch)(n, batchId, { ranges }, now);
-            n = advanceMemberToValidated(n, memberIssue, now);
-            return { state: n, result: undefined };
-        });
-        result.completed.push(unit(batchId));
-        advanceMemberOrValidate(deps, config, dispatch, batchId, batch.members.length, batch.executing_member, memberIssue, now, result);
+        completeMemberGate(deps, config, dispatch, batchId, batch, memberIssue, now, result);
         return;
     }
     if ((0, groundtruth_1.isMemberBlocked)(milestone) || dead) {
@@ -1065,11 +1279,17 @@ function reconcileMemberSlot(deps, config, dispatch, batchId, slot, now, result)
             : dead
                 ? 'agent-exited-unverified'
                 : 'member-blocked';
-        recordMemberRunLog(deps, dispatch, state0, batchId, batch, memberIssue, slot, now);
+        const lastTool = recordMemberRunLog(deps, dispatch, state0, batchId, batch, memberIssue, slot, now);
         journalEvent(deps, 'unit-failed', unit(batchId), {
             issue: memberIssue,
             reason,
             detail: 'member blocked',
+            // #591: attributes an unverified member exit to a concrete cause (e.g.
+            // `Monitor`) without opening the transcript. Gated on the RESOLVED reason, not
+            // `dead` — a member can be simultaneously `dead` AND carry a milestone-posted
+            // `reason` (it posted `blocked` and then exited), and that real block has no
+            // log-derived cause to attribute; only `agent-exited-unverified` does.
+            ...(reason === 'agent-exited-unverified' && lastTool !== null ? { last_tool: lastTool } : {}),
         });
         deps.store.withLock((s) => ({ state: releaseSlot(s, batchId, now), result: undefined }));
         // A member that never went green (RFC F.1) evicts DIRECTLY — no aggregate